Classify listing photos by renovation effort: no work, light work, heavy work, or CGI render.
The product was making about 50,000 vision-model calls per month to analyze property listing photos, costing about $3,000 per month ($36,000 per year). After the audit, the same workflow ran on a smaller, cheaper vision model with comparable accuracy, dropping the bill to roughly $200 per month ($2,400 per year). Net annual savings: $33,600. We accounted for tricky edge cases in the evaluation and provided a primary model plus fallback options for provider errors, rate limits, or outages.
Edge cases the audit specifically planned for
- Subtle CGI renders that look photographic at first glance
- Mixed listings (some real photos, some renders in the same gallery)
- Light-work vs major-work properties (fresh paint vs gut renovation)
- Construction-progress photos vs finished interiors
- Lighting, photo quality, and angle variance across providers
Vision tasks are easy to picture, but the same playbook applies to text classification, document extraction, ticket routing, RAG grading, content moderation, and more. Edge-case planning is the audit's value-add. Without it, score variance across models collapses and recommendations become unreliable.